METHOD
InKH converts user, market, portfolio, and tool events into structured operational knowledge via four components: passive knowledge injection for bounded working context, temporal graph memory for low-latency retrieval, a wiki audit surface for human-readable governance, and background extraction with maturity/decay/write-time invalidation.
KEY RESULT
46,080 baseline-conditioned evaluations across 24 seeds, 4 rounds, 80 episodes, 6 baselines. Mean task quality 0.815 at 900ms latency. −82.95% latency, −82.29% token cost, −96.58% stale-knowledge usage vs agent-driven wiki-walk memory, with +0.108 quality and +0.461 traceability improvement.
The most directly financial-AI paper in 15 sessions. Every Big 4 bank deploying agents for market analysis or trade preparation faces the same degradation pattern InKH solves. The wiki audit surface maps to APRA CPS 230. The design thesis — "adoption happens when complexity is absorbed by the system rather than transferred to the user" — is the enterprise deployment thesis for 2026-2027.
METHOD
First benchmark evaluating agent safety from final environment state after action sequences, not binary refusal checks. Categorizes violations by cause, enabling model-specific safety profiles. Places models in realistic agent-style projects with stateful workspaces.
KEY RESULT
Even the best-performing model has >54% harmful safety-violation rate (HSR). Current alignment remains fundamentally insufficient for realistic project environments where actions accumulate state. Different models exhibit distinct safety profiles.
Extends Safety Foundations Cracking to Layer 7: stateful workspace safety. Banking coding agents operate in exactly these stateful environments. An agent that passes single-turn safety checks may silently corrupt a project over multiple actions — the APRA CPS 230 operational resilience nightmare.
METHOD
AURA (Anonymization with Utility-Retention Adaptation) — a mask-reconstruct framework that decouples privacy localization from utility-preserving reconstruction. Adversarial privacy + utility-retention checks select candidates. Evaluated on real-user interview transcripts against web-search agent re-identification attacks.
KEY RESULT
Improves the privacy-utility frontier by using adaptive privacy scope to resist agentic re-identification. Mask-reconstruct method better preserves contextual utility under fixed privacy scope. 32 pages, 7 figures.
Agentic LLMs with web search fundamentally change the threat model for text anonymization. Banking: customer interaction transcripts, internal communications, regulatory submissions — all potentially re-identifiable by agentic adversaries even after standard anonymization. The mask-reconstruct approach provides a new privacy primitive for enterprise deployment.
Stateful Agent Safety (Layer 7) — SABER + AURA extend the Safety Foundations Cracking narrative. The concern has shifted from "can single-turn agents be safe?" to "at which of the 7+ layers will your specific multi-turn, stateful deployment break first?"
Inference Efficiency Renaissance continues — CLSA (7.6× decoding, 17.1× throughput at 128K) + SlidingServe (+30% capacity, −53% SLO violations). Combined with last week's Latent Reasoning Renaissance, the message is clear: inference cost is the #1 engineering problem and the field is attacking it from every angle.
Financial AI Infrastructure maturing — InKH is the second paper in 3 days directly addressing financial AI deployment. Combined with Shadow Price of Reasoning and Economy of Minds, a coherent subfield is forming: financial AI that is architecturally grounded, economically principled, and production-ready.
Economics of AI emerges as design paradigm — Economy of Minds (Hayekian coordination) → Shadow Price of Reasoning (marginal utility budget allocation) → InKH (complexity absorption). Economic formalism applied to AI at three levels.
Safety Foundations Cracking reaches 7 layers — The month opened with 2× 19/20 papers and is now shifting from diagnosis to remediation (Pre-Deployment Assurance, Recuse Signal, BraveGuard).
Latent Reasoning Renaissance stabilizes — 4 mechanisms in 5 days. The combined thesis — explicit token-space reasoning can be 40-60% cheaper — is now consensus.